A Data-Driven Hydrate Plug Detection in Offshore Gas & Condensate Flowlines

In multiphase transport pipelines, gas hydrate blockages are a major flow assurance challenge, due to the difficulties in remedial actions and the potential massive production losses. Scope of this study is to define an innovative data-driven methodology to early detect hydrates formation, provide an alarm, and permit early intervention before the complete blockage of the flowline. The proposed approach is applied on a Gas & Condensate pipeline during cold restart, which is the most critical scenario for hydrates formation. In the absence of reliable field data, the methodology was validated on synthetic data. Through a Design of Experiment (DoE) strategy, a wide range of operating conditions (with and without hydrate plug) has been simulated using a multiphase flow model, by varying some key parameters. Only field-measurable variables have been considered for the machine learning model training. In addition, a custom "Friction Factor" indicator and its derivative over time have been calculated, as they emerged to be crucial for model's performance enhancing. A classification model (XGBoost), called "Alarm Model", was defined to detect the formation of a hydrate plug and raise an alarm, based on a "RiskProbability" estimate. As a result, the model consistently managed to detect in advance the formation of hydrate plugs, particularly in cases of long-time formation. A second model (XGBoost), called "Failure Temporal Distance Model", was developed to classify the system status after an alarm is raised, giving information about the residual time to reach the failure event. This model showed good performances, with 85% of recall for the most critical class (i.e., imminent events), and a global accuracy of 80%. As a conclusion, this research highlights the successful application of machine learning and the relevance of the "Friction Factor" derivative in dynamically detecting plug formation in pipeline systems, without relying only on static thresholds. The use of DoE methodology has proven to be useful in obtaining sufficiently diverse simulations to achieve an algorithm that provides accurate and timely predictions. These findings contribute to the advancement of plug formation detection techniques, with potential applications in enhancing the operational efficiency and maintenance strategies of pipeline networks.